Sun Hao, Xiong Linghu, Huang Yi, Chen Xinkai, Yu Yongjian, Ye Shaozhen, Dong Hui, Jia Yuan, Zhang Wenwei
School of Mechanical Engineering and Automation, Fuzhou University, Fuzhou 350116, China.
Provincial Clinical College, Fujian Medical University, Fuzhou 350001, China.
Fundam Res. 2021 Dec 28;2(3):476-486. doi: 10.1016/j.fmre.2021.12.005. eCollection 2022 May.
Global pandemics such as COVID-19 have resulted in significant global social and economic disruption. Although polymerase chain reaction (PCR) is recommended as the standard test for identifying the SARS-CoV-2, conventional assays are time-consuming. In parallel, although artificial intelligence (AI) has been employed to contain the disease, the implementation of AI in PCR analytics, which may enhance the cognition of diagnostics, is quite rare. The information that the amplification curve reveals can reflect the dynamics of reactions. Here, we present a novel AI-aided on-chip approach by integrating deep learning with microfluidic paper-based analytical devices (µPADs) to detect synthetic RNA templates of the SARS-CoV-2 ORF1ab gene. The µPADs feature a multilayer structure by which the devices are compatible with conventional PCR instruments. During analysis, real-time PCR data were synchronously fed to three unsupervised learning models with deep neural networks, including RNN, LSTM, and GRU. Of these, the GRU is found to be most effective and accurate. Based on the experimentally obtained datasets, qualitative forecasting can be made as early as 13 cycles, which significantly enhances the efficiency of the PCR tests by 67.5% (∼40 min). Also, an accurate prediction of the end-point value of PCR curves can be obtained by GRU around 20 cycles. To further improve PCR testing efficiency, we also propose AI-aided dynamic evaluation criteria for determining critical cycle numbers, which enables real-time quantitative analysis of PCR tests. The presented approach is the first to integrate AI for on-chip PCR data analysis. It is capable of forecasting the final output and the trend of qPCR in addition to the conventional end-point Cq calculation. It is also capable of fully exploring the dynamics and intrinsic features of each reaction. This work leverages methodologies from diverse disciplines to provide perspectives and insights beyond the scope of a single scientific field. It is universally applicable and can be extended to multiple areas of fundamental research.
诸如新冠疫情这样的全球大流行已导致全球社会和经济严重混乱。尽管聚合酶链反应（PCR）被推荐为鉴定新冠病毒的标准检测方法，但传统检测耗时较长。与此同时，虽然人工智能（AI）已被用于控制疫情，但在PCR分析中应用AI以增强诊断认知的情况却十分罕见。扩增曲线所揭示的信息能够反映反应动态。在此，我们提出一种新型的人工智能辅助芯片方法，将深度学习与基于微流控纸的分析装置（µPADs）相结合，以检测新冠病毒ORF1ab基因的合成RNA模板。µPADs具有多层结构，使其能与传统PCR仪器兼容。分析过程中，实时PCR数据被同步输入到三个带有深度神经网络的无监督学习模型中，包括循环神经网络（RNN）、长短期记忆网络（LSTM）和门控循环单元（GRU）。其中，GRU被发现最为有效和准确。基于实验获得的数据集，早在第13个循环就能进行定性预测，这将PCR检测效率显著提高了67.5%（约40分钟）。此外，GRU在约第20个循环时能够准确预测PCR曲线的终点值。为进一步提高PCR检测效率，我们还提出了用于确定关键循环数的人工智能辅助动态评估标准，从而实现PCR检测的实时定量分析。所提出的方法首次将人工智能集成用于芯片PCR数据分析。除了传统的终点Cq计算外，它还能够预测qPCR的最终输出和趋势。它也能够充分探索每个反应的动态和内在特征。这项工作利用了不同学科的方法，提供了超越单一科学领域范围的观点和见解。它具有普遍适用性，可扩展到多个基础研究领域。
